A curious PM's experiments with AI, analytics, and building things.
The prototype worked, and the most valuable thing it produced wasn't the product - it was a list of things I'd do differently. Here's what I learned, and why I'm retooling.
Most people use Claude Code like a careful pair programmer. The bigger unlock is the opposite mindset - push it hard, run it in parallel, and build a roll cage so you don't crash. Here's the setup I've landed on for Trevorwithdata.
Funnels look like a solved problem until you try to build one for a journey analytics tool. Cross-device users, out-of-order events, conversion windows, person vs session scope - here's how we built a funnel block that handles all of it on top of a single ClickHouse function.
Setup is one thing. Actually exploring your data is another. Here's how we built notebooks as a canvas for reusable 'blocks' that can live inside a notebook or get dropped into an AI chat - starting with the most important block of all: the table builder.
Setting up a journey analytics platform means connecting warehouses, configuring datasets, building identity stitching, and defining metrics, dimensions, and segments - easily hours of work. Here's how we built an AI assistant that handles all of it through natural conversation, powered by the Claude Agent SDK and MCP tools.
Your analytics platform can count page views and sum revenue, but the metrics that actually drive decisions - conversion rate, average order value, revenue per visitor - require combining those building blocks with math that runs after aggregation. Here's how we built a calculated metrics engine with formula-level segments, statistical functions, and a live preview.
Traditional BI tools let you filter by 'who bought something.' Journey analytics segments let you filter by 'who browsed, then added to cart within 5 minutes, then purchased within the same session.' Here's how to build a segment engine with sequence awareness, scope control, and an AI builder that writes the segment logic for you.
Your events table has millions of rows, but an analyst just wants 'Revenue by Marketing Channel.' Here's how we built a semantic layer with dimension persistence and metric deduplication that turns messy data into clean, trustworthy reports - no window functions required.
Journey analytics requires sessionization, computed dimensions, segments, and smart aggregation - all at query time. Here's how to build a 6-layer SQL architecture to make it work.
A cookie on a laptop, a device ID on a phone, a CRM record in-store, a phone number at the call center - all the same person. How do you stitch that into one coherent journey? Here's how we built cross-device identity resolution for our DIY journey analytics platform.
When your data lives in separate tables with different schemas, how do you combine it all into one unified set of data for analysis? Here's how to build a 'Data Group' layer using ClickHouse materialized views.
A deep dive into how to use Airbyte to sync data from a data warehouse into ClickHouse - covering CDC, full refresh, and append-only sync modes, the tradeoffs of each, and how to build an API layer on top to keep things simple.
Before choosing any technology, I needed to establish what I would and would not build myself - and why that matters.
What if one PM with AI coding tools could build what usually takes hundreds of engineers? I'm finding out - and sharing everything I learn along the way.